Cyclic 2.5D Perceptual Loss for Cross-Modal 3D Medical Image Synthesis: T1w MRI to Tau PET

A team of researchers has published a significant advance in medical AI, introducing a novel 'Cyclic 2.5D Perceptual Loss' method for synthesizing 3D Tau Positron Emission Tomography (PET) scans from standard T1-weighted Magnetic Resonance Imaging (MRI) data. Published in *Human Brain Mapping*, the work addresses a major clinical barrier: PET imaging, crucial for diagnosing Alzheimer's disease via biomarkers like tau protein, is expensive, involves radioactive tracers, and has regulatory hurdles. The new AI framework reconstructs these valuable 3D pseudo-[18F]flortaucipir standardized uptake value ratio (SUVR) maps from routine structural MRI, making critical diagnostic information more accessible.

The technical innovation lies in the 'cyclic' training approach. Instead of using standard 2D, 3D, or static 2.5D perceptual losses, the model alternates optimization across the axial, coronal, and sagittal planes during training. This improves the volumetric consistency of the synthesized 3D PET images. Furthermore, the team standardized PET SUVRs by scanner manufacturer, a key step that reduces inter-manufacturer variability and better preserves high-uptake regions critical for assessing tau pathology. The method proved versatile, showing strong performance across multiple AI architectures including U-Net, UNETR, SwinUNETR, CycleGAN, and Pix2Pix, and was validated on data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and SCAN cohorts.

The result is a system that improves agreement between synthesized and real PET scans, particularly in brain regions relevant to Alzheimer's disease. By generating a synthetic biomarker readout from a common MRI scan, this technology could lower costs, reduce patient burden, and expand access to advanced neurological assessments. The code is publicly available, encouraging further development and validation in clinical settings.